Pipeline support for loading from deployment directories with config.json

[dbogunowicz]

PR description

This PR has two main goals:

1. Make pipelines read config.json file from deployment directory
2. Make pipelines accept deployment directory as a model_path argument.

Detailed changes:

• Refactor pipeline abstract method setup_onnx_file_path to setup_from_model, so it signals to the user that it does more things (e.g loads config.json file).
• Refactor utility function model_to_path to model_to_path_and_config, so it additionally supports loading config.json file.
• For vision pipelines add a setter for the class_names variable. This enables us to make sure that if the class_names are set up from the config.json, the user will be warned if they attempt to overwrite it.
• Make pipelines accept deployment directories (Directory class instance) as a model_path argument.

Manual testing plan

Make sure that vision pipelines accept deployment directory as well as Model as input and properly read class_names (mapping from labels to class names, e.g. '11' -> 'goldfish') from config.json:

## Manual Testing
import json
import os
import shutil

from deepsparse import Pipeline
from deepsparse.pipelines.helpers import DeploymentFiles
from sparsezoo import Model


def create_config_file_in_directory(deployment_dir_path, no_classes):
    dummy_class = "dummy_class"
    config_dict = {}
    config_dict[DeploymentFiles.ONNX_MODEL_FILE] = {label: dummy_class for label in range(no_classes)}
    with open(os.path.join(deployment_dir_path, DeploymentFiles.CONFIG_FILE), "w") as w:
        json.dump(config_dict, w)


def manual_testing(stub, num_classes, task_name):
    model = Model(stub)

    # create sample input
    pipeline_input = model.sample_batch(batch_size=1)['sample_inputs']
    args = {'images': pipeline_input[0]}

    # add mock config file to deployment directory
    path = model._path
    create_config_file_in_directory(model.deployment.path, num_classes)

    # create model with deployment directory...
    model = Model(path)

    # ...and feed it to the model
    deployment = model.deployment
    # pipeline = Pipeline.create(task_name, model) also works
    pipeline = Pipeline.create(task_name, deployment)

    out = pipeline(**args)
    if task_name == "image_classification":
        assert out.labels[0] == "dummy_class"
    elif task_name == "yolact":
        assert out.classes[0][0] == "dummy_class"
    else:
        # yolo
        assert out.labels[0][0] == "dummy_class"

    assert list(pipeline.class_names.keys())[0] == '0'
    assert list(pipeline.class_names.values())[0] == "dummy_class"


test_cases = {
    "image_classification": ("zoo:cv/classification/resnet_v1-50/pytorch/sparseml/imagenet/pruned95_quant-none", 1000),
    "yolact": ("zoo:cv/segmentation/yolact-darknet53/pytorch/dbolya/coco/pruned82_quant-none", 80),
    "yolo": ("zoo:cv/detection/yolov5-s/pytorch/ultralytics/coco/pruned_quant-aggressive_94", 80)
}

for task_name, (stub, num_classes) in test_cases.items():
    manual_testing(stub, num_classes, task_name)
print('done')

Testing annotate pipelines

Successful manual tests for yolo, yolact and IC.